MemorySync

Thanks, Adam. I agree; the database filter is the easy part. The harder part is anything that lives between requests, like a summary or a cached plan made for one user and then reused for the next. In MemorySync, every call has to name the end user, and a search only looks at that user's memories. That helps, but it can't protect a summary your agent keeps in its own working memory. Have you seen this happen in a real system, or mostly in testing?

When an agent codes in Cursor, there are two distinct context needs: Local repository state: Handled locally by Cursor’s codebase indexing, linters, and composer looking at your branch/lockfile. External platform contracts & API documentation: This is where models hallucinate deprecated endpoints, renamed parameters, or deleted SDK methods because the training cutoff is outdated. The Docs MCP specifically solves the second problem: giving the agent a live, zero-hallucination ground truth of the production API/SDK specs on demand, without polluting your prompt window with 500 lines of static markdown. Local code stays local; documentation stays fresh. Appreciate you reading and sharing this thought!